The calibration for many-body dissipative particle dynamics by using back-propagation neural networks

Abstract

ABSTRACT The many-body dissipative particle dynamics (MDPD) is a promising mesoscopic method for solving various interfacial issues. However, an accurate mapping between simulation parameters and corresponding fluid properties is still ambiguous in MDPD due to its additional many-body interaction. In the present paper, we adopt the back-propagation Neural Network (BPNN) model with high-fidelity database as a rigorous calibration approach for MDPD. The relations between three main, simulation parameters (i.e. attractive and repulsive coefficient A,B and cut-off radius of repulsive force ) and four real-fluid static and dynamic properties (i.e. density, viscosity, self-diffusion coefficient and surface tension) are achieved accurately with the frame of BPNN. The influences of hyperparameters and training sets on mean squared errors (MSE) are also provided. Moreover, in order to improve the training efficiency of the model, we put forward three data selection methods: random data selection, extreme value data selection and the orthogonal experiment data selection. We compare three sampling methods for training data, and the orthogonal experiment design data selection is proved to be the most effective way to reduce MSE. The BPNN model is the potential to be a powerful tool for determining the simulation parameters within given fluid properties accurately and efficiently in the MDPD system.